Deploying large deep learning models, especially object recognition models like YOLOv11n, on resource-constrained CPU platforms is always a big challenge in achieving high computational efficiency and fast inference speed. In particular, CPUs that are not equipped with dedicated acceleration hardware such as GPUs will make processing complex models costly. To address this issue, this research proposed a method that focuses on optimizing the performance of the YOLOv11n model to enhance it efficiently on general-purpose CPUs without using dedicated hardware. In this research, a novel technique is introduced to reduce the model size and speed up inference without significantly reducing the accuracy of the model using the method of weight quantization and selective activations. The experimental analysis results of this study have demonstrated the correctness and effectiveness by quantizing the weights and keeping the sensitive activations in FP32 format to maintain the stability and accuracy of the model while minimizing the consumption of processor resources.